Learned Multiagent Real-Time Guidance with Applications to Quadrotor Runway Inspection
Résumé
Aircraft runways are periodically inspected for debris and damage. Instead of having pilots coordinate the motion of the quadrotors manually or hand-crafting the desired quadrotor behavior into a guidance law, this paper reports the use of deep reinforcement learning to learn a closed-loop multiagent real-time guidance strategy for quadrotors to autonomously perform such inspections. This yields a significant reduction in engineering effort while enabling highly-flexible real-time performance. The runway is discretized into a number of rectangular tiles, which must all be visited for the runway to be considered inspected. The guidance system reported here calculates a desired acceleration in real time for the quadrotor(s) to track in order to complete the task. This paper first develops the guidance technique, trains it in simulation, and evaluates it experimentally using an indoor quadrotor laboratory. This process is then repeated for an outdoor setting on a real runway, where the proposed guidance strategy is compared to a handcrafted strategy and applied to a multiquadrotor scenario where the quadrotors must learn to coordinate their behavior and be resilient to the failure of one quadrotor mid-experiment. Multiagent, fault-tolerant, learned behavior is successfully demonstrated through outdoor quadrotor flights. Additional simulations and experiments demonstrate the technique is viable in a swarm with additional quadrotors, on a variety of runway shapes and with increased discretization of the runway. This work shows how modern learning-based techniques can: 1) reduce the engineering effort required to design complex guidance systems and 2) be implemented on real hardware in a representative outdoor environment.
Origine | Fichiers éditeurs autorisés sur une archive ouverte |
---|